Four-step safety process for AI models like Gemini
Google DeepMind and Isomorphic Labs say they follow four steps to ensure their models, such as Gemini, are safe and useful to experts: threat modeling, evaluations, mitigations and monitoring.
Supporting evidence
Google DeepMind and Isomorphic Labs approach to bioresilience
Original excerpt
To ensure our models, like Gemini, are safe and useful to experts, we follow a four-step safety process: threat modeling, evaluations, mitigations and monitoring.